Video analytics is the automated extraction of meaningful information from video — software that watches the footage and reports what is happening, instead of leaving it to a human to notice. It spans a family of capabilities built on one foundation (detecting objects in a frame): classification, tracking and re-identification, face recognition, licence-plate reading, behavioural rules, anomaly detection, and forensic search. Each answers a different question, from "is there a person here?" to "find everyone who entered after 6 pm".
In a surveillance system, analytics is what turns passive recording into something proactive and searchable. Results surface as metadata over ONVIF Profile M, so the VMS can alert on them live, draw them on the video, and index them for later search. The work can run on the camera (edge), an on-prem server, or the cloud, and that placement decides latency and cost.
The defining discipline is honesty about accuracy. Every analytic has a realistic precision/recall range that depends on scene, lighting, angle, and tuning — never a single "99%" and never 100%. Two analytics deserve a legal flag before any technical discussion: face recognition and licence-plate reading produce data about identifiable people and trigger privacy law (GDPR, BIPA). The model internals belong to the AI for Video Engineering section; this section covers how each analytic plugs into the camera, the VMS, and storage.
